How do we become confident in the safety of a machine learning system? — AI Alignment Forum